Problem
Draft PR #38 reduces the Windows ARM64 Vulkan build from roughly 33 minutes to about 8 minutes by cross-building with Clang/Ninja and optimizing the host-side Vulkan target for size. Packaging and ABI checks passed at its historical head, but no Windows ARM64 Vulkan device validated runtime correctness or performance, and the branch is now conflicted with current release automation.
Acceptance criteria
- integrate current native main without force-pushing and reconcile the current release workflow;
- rebuild the full Windows ARM64 Vulkan artifact on the exact PR head;
- validate model load and representative generation on physical Windows ARM64 Vulkan hardware;
- compare output correctness, stability, and representative throughput against the current release build;
- confirm PE architecture, DLL inventory/imports, wrapper exports, KleidiAI, and package provenance;
- keep release/publication as a separate explicit approval boundary.
Existing work
Draft PR #38 contains the build configuration and historical non-publishing build evidence. Do not merge it until the device qualification is complete.
Problem
Draft PR #38 reduces the Windows ARM64 Vulkan build from roughly 33 minutes to about 8 minutes by cross-building with Clang/Ninja and optimizing the host-side Vulkan target for size. Packaging and ABI checks passed at its historical head, but no Windows ARM64 Vulkan device validated runtime correctness or performance, and the branch is now conflicted with current release automation.
Acceptance criteria
Existing work
Draft PR #38 contains the build configuration and historical non-publishing build evidence. Do not merge it until the device qualification is complete.